Hybrid Evolution of Heterogeneous Neural Networks

In this paper we are describing experiments and results of applications of the continual evolution algorithm to construction and optimization of recurrent neural networks with hetrogeneous units. Our algorithm is a hybrid genetic algorithm with sequential individuals replacement, varibale population size and age-based probability control functions. Short introduction to main idea of the algorithm is given. We describe some new features implemented into the algorithm, the encoding of individuals, crossover, and mutation operators. The behavior of population during an evolutionary process is studied on atificial benchmark data sets. Results of the experiments confirm the theoretical properties of the algorithm.